[Soc-2018-dev] Weekly report #11 - Implementing a Hair Shader for Cycles

[Leonardo E. Segovia]

Week 11

Hi everyone! This is my report for the present week.
Objectives

Brecht requested that I profile the shader to see if there were any
possible optimizations. Also, if I could find out and fix why the shader
was noisy in certain areas.
What's been done

   - Brecht uploaded the missing regression tests (rBL62073
   <http://developer.blender.org/rBL62073>).
   - Apply Filter Glossy to the Roughness value (as detailed in the Blender
   manual), rather than the corresponding logistic's variance (
   rB52a0e67fe0050683a888d6643ec2035bd2b3ea1c
   <http://developer.blender.org/rB52a0e67fe0050683a888d6643ec2035bd2b3ea1c>
   ).
   - Enable mode weights. This lets you select which light mode will be
   present in the render. Not exactly useful for physically-based renders, but
   for debugging issues like noise. (
   rB2ab98cca19cb3acd3da61b5f71a61a1139f1713e
   <http://developer.blender.org/rB2ab98cca19cb3acd3da61b5f71a61a1139f1713e>,
   rB21aefa098f4935597a759e199153e9b55279ff8a
   <http://developer.blender.org/rB21aefa098f4935597a759e199153e9b55279ff8a>
   )

I profiled the shader using nirved's Hair Curls (4 curls, 1000 hairs per
curl, black background, lighted by a HDR sphere) and Xcode's Instruments.
The only edit was to make *all* functions in bsdf_hair_principled.h
ccl_device (ie. not inlined).

Top ten of functions (full stacktrace as an addendum to the previous week's
report in soc-2018-dev):

9.11 min   32.6%    9.11 min
ccl::QBVH_bvh_intersect_shadow_all_hair(ccl::KernelGlobals*, ccl::Ray
const*, ccl::Intersection*, unsigned int, unsigned int, unsigned int*)
6.28 min   22.5%    6.28 min
ccl::QBVH_bvh_intersect_hair(ccl::KernelGlobals*, ccl::Ray const*,
ccl::Intersection*, unsigned int, unsigned int*, float, float)
56.99 s    3.4% 56.99 s
ccl::BVHObjectBinning::BVHObjectBinning(ccl::BVHRange const&,
ccl::BVHReference*, ccl::BVHUnaligned const*, ccl::Transform const*)
48.82 s    2.9% 48.82 s
ccl::kernel_branched_path_trace(ccl::KernelGlobals*, float*, int, int,
int, int, int)
36.48 s    2.1% 36.48 s
ccl::camera_sample_perspective(ccl::KernelGlobals*, float, float,
float, float, ccl::Ray*)
29.71 s    1.7% 29.71 s
ccl::kernel_branched_path_integrate(ccl::KernelGlobals*, unsigned int,
int, ccl::Ray, float*, ccl::PathRadiance*)
29.22 s    1.7% 29.22 s
ccl::bsdf_principled_hair_eval(ccl::KernelGlobals*, ccl::ShaderData
const*, ccl::ShaderClosure const*, ccl::float3, float*)
21.94 s    1.3% 21.94 s
ccl::kernel_tex_image_interp(ccl::KernelGlobals*, int, float, float)
21.39 s    1.2% 21.39 s      ccl::svm_eval_nodes(ccl::KernelGlobals*,
ccl::ShaderData*, ccl::PathState*, ccl::ShaderType, int)
20.40 s    1.2% 20.40 s
ccl::svm_node_mapping(ccl::KernelGlobals*, ccl::ShaderData*, float*,
unsigned int, unsigned int, int*)

Even though bsdf_principled_hair_eval appears (as expected, since it's
pretty heavy in calculations), the overwhelming majority of CPU time goes
to the BVH intersection calculation for hair, so I deemed this to be
sufficiently optimized.

As regards noise, I could only find out where it happens and in which
illumination conditions, not the actual source (code) of the noise.
<https://wiki.blender.org/wiki/File:Hair-Curls-noise.png>
Transparent render of nirved's Hair Curls. From left to right: full render,
only primary specular (R), only transmission (TT), only secondary specular
(TRT), only residual (TRRT+). Noise is most prevalent at the extrema of the
curl (which is away from the light).

   - Noise is most visible in the penumbra of the hair (places away from
   the glint), but separating the light passes (below) shows it affects the
   whole length of the hair.
   - It arises in the Gloss Ind (i.e. indirect glossy) pass of the
   renderer, which AFAIK points to light reflection between hairs.

<https://wiki.blender.org/wiki/File:Hair-Curls-noise.glossind.png>
Indirect Glossy pass of nirved's Hair Curls. Noise is present throughout
the hair length.

   - It only affects the TT mode, which is consistent with the observations
   since it alters the coloring of the render. Rendering separate layers for
   each mode shows the rest are unaffected.

Next up

   - Nothing yet, I'll wait to hear back from Lukas/Brecht to see if they
   have any more ideas/advice for debugging the noise issue.

QuestionsNone this time.

-- 
Lic. Leonardo E. Segovia
Departamento de Ciencias e Ingeniería de la Computación
Universidad Nacional del Sur
San Andrés 800 - Campus Palihue, B8000 Bahía Blanca, Argentina
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.blender.org/pipermail/soc-2018-dev/attachments/20180727/d05de30a/attachment.html>